2026-07-15 –, Memorial Hall
Field inspections in agriculture and science face a common problem: hands are full, data needs structure, and decisions require both inspection history and domain expertise. I built HiveGuide, an open-source field inspection system with three main components: (1) voice transcription for hands-free data entry, (2) AI extraction to structured data and action items, and (3) an AI assistant that provides intelligent advice by routing between personal inspection history and authoritative domain literature. For the assistant, I tested 7 routing strategies on 500+ queries to solve a dual-source problem: when to query your data versus domain references. The LLM classifier approach balanced accuracy and speed without requiring training data. The architecture is transferable to any inspection domain where you need minimal device interaction and intelligent advising.
Background
Field work across agriculture, environmental monitoring, and scientific research shares common constraints: practitioners need structured data capture while their hands are occupied, dirty, or gloved. Current solutions—voice memos, note apps, or digitized forms—produce unstructured data that's difficult to query or analyze. More critically, field decisions require combining two distinct knowledge sources: personal inspection history ("what's normal for my sites?") and authoritative domain knowledge ("what do experts recommend?"). Generic AI assistants can't access your data; domain-specific apps don't leverage expert knowledge.
I developed HiveGuide to solve this for beekeeping inspections, where you're holding frames with thousands of stinging insects while wearing propolis-covered gloves. The architecture proved generalizable to any inspection workflow requiring minimal interaction, structured capture, and intelligent advising.
Methods
The system is built on a Python FastAPI backend with PostgreSQL database, using LangChain for the RAG architecture and OpenAI APIs for transcription and language models. The frontend is React Native (iOS) and React Native Web, with the Python backend handling all AI/ML processing.
The system has three components:
- Voice Transcription: Real-time streaming transcription (2-second delay) via iOS native app. Audio sent to server in chunks every few seconds, minimizing data loss risk compared to batch processing. Platform choice (native vs web) drove capability—web apps can't achieve this latency or reliability.
- AI Extraction: LLM converts voice notes to structured fields. Example: "It's in the 60s and cloudy. Fresh eggs in good pattern, didn't spot the queen" extracts temperature, queen_visible: False, eggs_visible: True, laying_pattern: "solid". Structured data enables querying and generates automated action items based on inspection findings.
AI Assistant with Intelligent Routing: This solved the core technical problem. Field inspection questions require either personal data ("Is my hive at normal weight?"), domain knowledge ("What causes bee dysentery?"), or both ("Is my hive's October weight normal for Wisconsin?"). I implemented 7 routing approaches:
- LLM classifier (pre-classifies query intent)
- Heuristic rules (keyword matching)
- Embedding similarity (query vector vs source vectors)
- Supervised classifier (trained on labeled queries)
- Agent-based (agent decides tool usage)
- Hybrid combinations
- Always-both baseline
Each routed to SQL database (inspection history) and/or vector search with pgvector (RAG over authoritative sources). A LangChain agent synthesized retrieved context. Validation layer caught generic responses and forced retry.
Results
Testing on 500+ queries:
- Supervised classifier: 97.8% accuracy (highest), requires labeled training data
- LLM classifier: 95.2% accuracy, ~1s overhead, no training needed—selected for deployment
- Agent-based: good retrieval, higher error rates from increased complexity
The LLM classifier balanced performance with practical deployment constraints. In use across multiple hives over several months, the system successfully generated structured inspection data, automated task lists, and provided contextualized advice combining personal history with domain references.
To mitigate hallucination risk, responses link directly to source materials with specific page citations.
Generalizability
This pattern applies wherever you need:
- Minimal device interaction (hands busy/dirty)
- Structured data for later analysis
- Decisions based on inspection history + domain expertise
Examples: equipment maintenance, scientific field inspections, beekeeping, etc.
Conclusion
The dual-source routing problem appears across scientific and agricultural field work but lacks established solutions. Systematic testing of routing strategies showed LLM classifiers provide practical performance without training overhead. The architecture is open-sourced (Creative Commons NC) for adaptation to other inspection domains.
Links
- GitHub: github.com/CarolynOlsen/hiveguide_public
- Medium writeup: https://medium.com/@carolyn.olsen/ai-powered-field-inspection-app-design-for-agriculture-and-science-a4507b85e30e
- arXiv pre-print on routing strategies comparison is upcoming
Carolyn Olsen developed HiveGuide as an independent open-source project to solve practical field inspection challenges in beekeeping and generalized for broader scientific applications. In her day role, she is Director of Data Science at The Hartford, where she leads an AI Accelerator for two business areas, helping them leverage generative AI. Previously VP of Data Science at Clearcover, she has extensive experience developing production AI systems including LLM-powered tools, supervised ML models, and reinforcement learning models. She holds a Master of Science in Applied Economics from Marquette University and served 8 years in the U.S. Coast Guard Reserve.